Shading forecasting Interval in time series in R using ggplot

Question

Please refer to the dput of data . You may directly scroll down to objective and problem statement. Maybe you don't need data as you could have encountered this problem earlier.

Calling required libraries

library(zoo)
library(ggplot2)
library(scales)
library(plotly)
library(ggthemes)
library(forecast)
library(plotly)
library(DescTools)

dput of data

dput(ridership.ts)
structure(c(1709L, 1621L, 1973L, 1812L, 1975L, 1862L, 1940L, 
2013L, 1596L, 1725L, 1676L, 1814L, 1615L, 1557L, 1891L, 1956L, 
1885L, 1623L, 1903L, 1997L, 1704L, 1810L, 1862L, 1875L, 1705L, 
1619L, 1837L, 1957L, 1917L, 1882L, 1933L, 1996L, 1673L, 1753L, 
1720L, 1734L, 1563L, 1574L, 1903L, 1834L, 1831L, 1776L, 1868L, 
1907L, 1686L, 1779L, 1776L, 1783L, 1548L, 1497L, 1798L, 1733L, 
1772L, 1761L, 1792L, 1875L, 1571L, 1647L, 1673L, 1657L, 1382L, 
1361L, 1559L, 1608L, 1697L, 1693L, 1836L, 1943L, 1551L, 1687L, 
1576L, 1700L, 1397L, 1372L, 1708L, 1655L, 1763L, 1776L, 1934L, 
2008L, 1616L, 1774L, 1732L, 1797L, 1570L, 1413L, 1755L, 1825L, 
1843L, 1826L, 1968L, 1922L, 1670L, 1791L, 1817L, 1847L, 1599L, 
1549L, 1832L, 1840L, 1846L, 1865L, 1966L, 1949L, 1607L, 1804L, 
1850L, 1836L, 1542L, 1617L, 1920L, 1971L, 1992L, 2010L, 2054L, 
2097L, 1824L, 1977L, 1981L, 2000L, 1683L, 1663L, 2008L, 2024L, 
2047L, 2073L, 2127L, 2203L, 1708L, 1951L, 1974L, 1985L, 1760L, 
1771L, 2020L, 2048L, 2069L, 1994L, 2075L, 2027L, 1734L, 1917L, 
1858L, 1996L, 1778L, 1749L, 2066L, 2099L, 2105L, 2130L, 2223L, 
2174L, 1931L, 2121L, 2076L, 2141L, 1832L, 1838L, 2132L), .Tsp = c(1991, 
2004.16666666667, 12), class = "ts")

Creating data frame of ts object to use ggplot

tsd = data.frame(time = as.Date(ridership.ts), 
                 value = as.matrix(ridership.ts))

Building linear model

ridership.lm <- tslm(ridership.ts ~ trend + I(trend^2))

Adding new column to existing data frame tsd

tsd$linear_fit = as.matrix(ridership.lm$fitted.values)

Defining length of validation and training period

nValid = 36 
nTrain = length(ridership.ts) - nValid

Training data

train.ts = window(ridership.ts, 
                  start = c(1991, 1),
                  end = c(1991, nTrain))

validation data

valid.ts = window(ridership.ts, 
                  start = c(1991, nTrain + 1), 
                  end = c(1991, nTrain + nValid))

Building model

ridership.lm = tslm(train.ts ~ trend + I(trend^2))

Forecasting using our build model

ridership.lm.pred = forecast(ridership.lm, h = nValid, level = 0)

Making dataframe for the fitted model values

tsd_train_model = data.frame(time = as.Date(train.ts), 
                             lm_fit_train = as.matrix(ridership.lm$fitted.values))

Making dataframe for plotting purpose

forecast_df = data.frame(time = as.Date(valid.ts), 
                         value = as.matrix(ridership.lm.pred$mean))

Creating plot using ggplot

p1 = ggplot(data = tsd, 
            aes(x = time, y = value)) + 
  geom_line(color = 'blue') + 
  ylim(1300, 2300) + 
  geom_line(data = tsd_train_model, 
            aes(x = time, y = lm_fit_train), 
            color = 'red')

p2 = p1 + 
  geom_line(data = forecast_df, 
            aes(x = time, y = value), 
            col = 'red', linetype = 'dotted') + 
  scale_x_date(breaks = date_breaks('1 years'), 
               labels = date_format('%b-%y')) +
  geom_vline(xintercept = as.numeric(c(tsd_train_model[NROW(tsd_train_model), ]$time,  #last date of training period
                                       forecast_df[NROW(forecast_df), ]$time))) #last date of testing period 

p3 = p2 + 
  annotate('text', 
           x = c(tsd_train_model[NROW(tsd_train_model)/2, ]$time, 
                 forecast_df[NROW(forecast_df) / 2,]$time),
           y = 2250, 
           label = c('Training Period', 'Validation Period'))

Objective: I want to add forecast error of 5 percentile and 95 percentile on both side of predicted line (dotted red in this figure) and shade the region.

I used quantile for producting forecast range

q = quantile(ridership.lm.pred$residuals, c(.05, .95))

percentile_5 = as.numeric(q[1])
percentile_95 = as.numeric(q[2])

Add 5 percentile and 95 percentile to the forecast data

yl = forecast_df$value + percentile_5 
ym =  forecast_df$value  + percentile_95

Problem: If I use the below command then it is not displaying the shaded region for the complete validation period.

p3 + geom_ribbon(data = forecast_df, 
                 aes(ymin = yl, 
                     ymax = ym), 
                 fill="gray30")

NROW(yl)
 [1]36

sum(is.na(yl))
[1] 0

NROW(ym)
[1] 36

sum(is.na(ym))
[1] 0

Things Tried: If I replace the value of ymin and ymax by any other value for example If I use the below command then I get the figure shown just below the command

p3 + geom_ribbon(data = forecast_df, 
                 aes(ymin = rep(1750,36), 
                     ymax = rep(2000,36), 
                     fill="gray30"))

My Questions:

Can anyone please tell me the reason behind the output in figure 2? Why R is giving strange output as in figure 2?

Can anyone please help me to shade the complete region using ggplot?

Please indicate all the packages you used in the code. I don't think `tslm` is part of the base packages. Same with conversion of `ts` object to `Date` object. This would help others reproduce your problem for troubleshooting. — Z.Lin, Oct 15 '17 at 03:31

Z.Lin · Accepted Answer · 2017-10-18T04:14:48.357

TLDR: remove the line ylim(1300, 2300) + from your ggplot code.

When you set your plot's limits using scale_x_***() / scale_y_*** (or equivalently xlim() / ylim()), the plot will throw away all data points that fall outside this range. In the case of a geom_ribbon that requires both ymin & ymax values, when the values corresponding to ymax get removed (since they are larger than 2300), the ribbon can't be plotted with only ymin, hence the ribbon stops short before then.

If you really want to plot only for the range (1300, 2300), set the limits inside coord_cartesian() instead. This enables the plot to zoom to the range limit, without discarding data points outside. See the documentation for more information.

Other non-essential suggestions below:

For plotting in ggplot, I'd usually try to keep everything within the same data frame, as far as possible, to utilise common variables in the aesthetic mappings. Here's how I'd do it:

Combining everything into a single data frame:

library(dplyr)
df <- left_join(tsd %>% select(time, value),
                rbind(tsd_train_model %>% 
                        rename(fit = lm_fit_train) %>%
                        mutate(status = "train"),
                      forecast_df %>%
                        rename(fit = value) %>%
                        mutate(status = "valid")))
df <- df %>%
  mutate(yl = ifelse(status == "valid", fit + percentile_5, NA),
         ym = ifelse(status == "valid", fit + percentile_95, NA))

> head(df)
        time value      fit status yl ym
1 1991-01-01  1709 1882.681  train NA NA
2 1991-02-01  1621 1876.546  train NA NA
3 1991-03-01  1973 1870.518  train NA NA
4 1991-04-01  1812 1864.597  train NA NA
5 1991-05-01  1975 1858.784  train NA NA
6 1991-06-01  1862 1853.078  train NA NA

> tail(df)
          time value      fit status       yl       ym
154 2003-10-01  2121 2190.490  valid 1934.914 2397.875
155 2003-11-01  2076 2200.756  valid 1945.179 2408.141
156 2003-12-01  2141 2211.129  valid 1955.553 2418.514
157 2004-01-01  1832 2221.609  valid 1966.033 2428.994
158 2004-02-01  1838 2232.197  valid 1976.620 2439.582
159 2004-03-01  2132 2242.891  valid 1987.315 2450.277

Plot

ggplot(data = df,
       aes(x = time)) +

  # place the ribbon below all other geoms for easier viewing, & increase transparency
  geom_ribbon(aes(ymin = yl, ymax = ym), fill = "gray30", alpha = 0.2) +

  # original values
  geom_line(aes(y = value), color = "blue") +

  # fitted values (line type differs by training / validation)
  geom_line(aes(y = fit, linetype = status), color = "red") +

  # indicates validation range
  geom_vline(xintercept = c(min(df$time[df$status=="valid"]),
                            max(df$time[df$status=="valid"]))) +

  scale_x_date(breaks = scales::date_breaks('1 year'), 
               labels = scales::date_format('%b-%y')) +

  # hide legend for line type (comment this line out if you want to show it)
  scale_linetype(guide = F) + 

  # limits can be tweaked here
  coord_cartesian(ylim = c(1300, 2500)) +

  # plain white plot background for easier viewing
  theme_classic()

Edit: Alternative solution that makes legends easier:

# create long data frame where all values (original / training / validation) are
# in the same column
df2 <- rbind(tsd %>% select(time, value) %>%
               mutate(status = "original"),
             tsd_train_model %>% 
               rename(value = lm_fit_train) %>%
               mutate(status = "train"),
             forecast_df %>%
               mutate(status = "valid")) %>%
  mutate(yl = ifelse(status == "valid", value + percentile_5, NA),
         ym = ifelse(status == "valid", value + percentile_95, NA))

# in the scales for colour / line type, define the same labels in order to
# combine the two legends
ggplot(data = df2,
       aes(x = time)) +
  geom_ribbon(data = subset(df2, !is.na(yl)),
              aes(ymin = yl, ymax = ym, fill = "interval"), alpha = 0.2) +
  geom_line(aes(y = value, color = status, linetype = status)) +
  geom_vline(xintercept = c(min(df2$time[df$status=="valid"]),
                            max(df2$time[df$status=="valid"]))) +
  scale_x_date(breaks = scales::date_breaks('1 year'), 
               labels = scales::date_format('%b-%y')) +
  scale_color_manual(name = "",
                     values = c("original" = "blue",
                                "train" = "red",
                                "valid" = "red")) +
  scale_linetype_manual(name = "",
                 values = c("original" = "solid",
                            "train" = "solid",
                            "valid" = "longdash")) +
  scale_fill_manual(name = "",
                    values = c("interval" = "gray30")) +
  coord_cartesian(ylim = c(1300, 2500)) +
  theme_classic() +
  theme(legend.position = "bottom")

can you please tell me how can I add legends to this graph both automatically and manually ? — learner, Oct 16 '17 at 11:30
what legend do you want? Colour (blue for original observations, red for fitted values) / line type (unbroken for training, dashed for validation) / error range? — Z.Lin, Oct 16 '17 at 13:29
Yes. You are correct. Can you please tell me how can I add those ? I tried scale_color_manual but this doesn't seems to be working . Please guide — learner, Oct 16 '17 at 16:05

Shading forecasting Interval in time series in R using ggplot

1 Answers1